How has the increase of Automated Trading Systems (ATS) influenced the futures market?

• CME transaction data identitifies ATS with the 1028 tag
• Data available from November 2012 to present
• Consider data form 2012 to 2016
• Use gradient plots to highlight the changes
• Graphs are interactive

MAN vs ATS - Group

MAN vs ATS - Agriculture

MAN vs ATS - Energy

Trade Participation

Where are the opportunities?

• Limited, Local, Spectrum
  • Less liquid parts of the curve
  • Changes to macro
  • Concentrated risk on asymmetric risk/reward trades

• Quantitative
  • Harvest systematic alpha from large universe of commodities
  • Rule based decision making process
  • Strategies inspired by discretional methodology

Literature

• Algorithmic trading and Market Dynamics

• Automated Trading in Futures Markets, Richard Haynes and John S. Roberts

• Automated Trading in Futures Markets - Update, Richard Haynes and John S. Roberts

Law of active portfolio management

\[ \text{Performance} = \text{Skill} \times \sqrt{\text{Breadth}} \]

Suppose we are in a coin flipping casino

• Flip coins in stead of futures
• The coin is biased - \(P(\text{heads}) = 0.51\)

How does the betting work?

• We have 1000 coins
• The minimum wager is 1 coin
• If you win you gain 1 coin
• If you loose you loose 1 coin
• There are 1000 tables with coin wagers
• Games runs in parallel

What is the optimal way to allocate coins?

Two extremes

• Bet 1000 coins on one coin flip
• Bet 1 coin on 1000 coin flips

Expected Return

• Single bet: \(0.51 \times 1000 + 0.49 \times (-1000) = 20\)

• Multi bet: \(1000 \times [0.51 + 0.49 \times (-1)] = 20\)

The same expected return

Risk - Probability to lose it all:

• Single bet: 49%

• Multi bet: \(0.49 \times 0.49 \times \dots \times 0.49 = 0.49^{1000} \approx 0\)

Risk - Standard Deviation:

• One coin per table

\[ \text{risk} := \text{std}\left\{1,-1,-1,1, \dots, 1 \right\} = 1 \]

• One 1000 coin bet, 999 zero coin bets

\[ \begin{align} \text{risk} &:= \text{std}\left\{1000,0,0,0, \dots, 0 \right\} = 31.62 \\ \text{risk} &:= \text{std}\left\{-1000,0,0,0, \dots, 0 \right\} = 31.62 \end{align} \]

Coin Flip Casino - Reward/Risk

• Just like Sharpe Ratio

• Single bet: \(\text{SR}_{\text{single}} = \frac{20}{31.62} =0.63\)

• Multi bet: \(\text{SR}_{\text{multiple}} = \frac{20}{1} =20\)

Coin flipping casino - Observation

• \(20 = 0.63 \times \sqrt{1000}\)

• \(\text{SR}_{\text{multiple}} = \text{SR}_{\text{single}} \times \sqrt{\text{Bets}}\)

• \(\text{Performance} = \text{Skill} \times \sqrt{\text{Breadth}}\)

How does this apply to commodity futures?

• We use insights gained from years of fundamental trading to inspire bespoke quantitative strategies that are applied to a large collection of commodity markets

• We increase breadth or diversification by
  • how,
  • what and
  • when we trade

When we trade

What we trade

Literature

• The fundamental law of active management - Richard C. Grinold
• Extending the Fundamental Law of Investment Management - JP Morgan US Equity Quantitative Research

• Continuous Futures Price Series
  • We require long time series data
  • Futures expire too soon to gather sufficient data
  • How do you handle rolls?

• Non-stationarity of Price Data
  • Time series data can only reliably be forecasted if stationary
  • Machine Learning algorithms are designed for stationary features
  • How do we create stationary data?

Continuous Futures Price Series

Continuous Futures Curves

Stationarity

Stationarity

How to obtain a stationary time series

• Traditional
  • Price differences
  • Returns
  • Memory loss

• Modern
  • Fractional differences
  • Memory present

Stationarity

Stationarity

Carry overview

• Take advantage of curve shape
• Contango is typically less steep in further dated contracts, i.e. it has a lower roll yield than near dated contracts
• Extract roll yield with a short position in the near and a long position in the far dated contracts of the same commodity

Carry Example

Carry Example

Flavours of carry

• Carry - inspired by discretionary methodology (BB1)
  • 30 commodities and 2 tenors
  • Sizing determined using percentile methodology
  • Risk 0.5% of strategy NAV per trade
  • Monthly rebalance

• Machine Learning Enhanced Carry (BB2)
  • 30 commodities and 2 tenors
  • Fractionally differenced time-series
  • Meta labelling
  • Ensemble machine learning
  • Deterine probability of profitable trade and size trade accordingly

Backtest Assumptions

• $10 Round trip of each contract
• Slippage of 1 tick per contract

• Entry cost of one spread \(= \text{\$}10 + 2 \text{ ticks} \times \text{multiplier}\)
• Exit cost of one spread \(= \text{\$}10 + 2 \text{ ticks} \times \text{multiplier}\)

BB1 - Statistics

BB1 - Statistics

BB1 - Statistics

BB2 - Statistics

BB2 - Statistics

BB2 - Statistics

Systematic Carry - Statistics

Carry literature

Literature on extracting carry from futures:

• An Anatomy of Commodity Futures Risk Premia - Szymanowska et al.
• Commodity convenience yield and risk premium determination - Wei and Zhu
• Carry - Koijen et al.
• The Strategic and Tactical Value of Commodity Futures - Erb and Harvey

Literature on applying machine learing techniques in algorithmic trading:

• Advanced in financial machine learning - Marcos López de Prado
• The 10 Reasons Most Machine Learning Funds Fail

Trend overview

• Trend following is about absolute performance of each commodity
• Identify trends over selection of time frames
• Slowly build position as trend increases
• Slowly exit position as trend decreases
• 35 Commodities and 2 tenors

Designing, not fitting a strategy

Aim of a trend following strategy

• Profitably trade trending markets
• Accross a diverse universe of commodities

If we feed our trend system fake trendy data

• linear and sinusoidal, with
• Gaussian noise

can we trade it profitably?

Fake Trendy Data

Trend - Fake Data Performance

Trend - Statistics

Trend - Statistics

Trend - Statistics

Trend - Statistics

Trend literature

• The Strategic and Tactical Value of Commodity Futures - Erb and Harvey
• The fundamentals of Commodity Futures Returns
• Time series momentum - Moskowitz et al.
• Which trend is your friend - Levine and Pedersen
• Trend Following with Managed Futures - Greyserman and Kaminski

Relative Roll overview

Strategy not yet live.

Relative Roll literature

• Pairs Trading: Performance of a Relative Value Arbitrage Rule - Gatev et al.
• Pairs Trading - Vidyamurthy
• The Application of Pairs Trading to Energy Futures Markets - Kanamura et al.

Sentiment overview

Strategy not yet live.

Sentiment literature

• Machine learning and event detection for trading energy and metals futures - Hafez et al.
• JP Morgan: Using Ravenpack sentiment for cross asset style timing - Hafez et al.
• Macro Fundamentals or Geopolitical Events? A Textual Analysis of News Events for Crude Oil - Brandt and Gao

Investment Thesis

• Focus on the intersection of technolody, data and behavioral finance applied to the broad commodity space
• Build strategies based on sound economic theory to help deliver long-term repeatable results. Inspired by
  • academic and
  • proprietary research.
• Investment process built on the scientific method consisting of the systematic
  • observation,
  • measurement,
  • experiment,
  • hypothesis formulation,
  • testing and
  • modification of hypothesis

Quantitative Portfolio - Statistics

Quantitative Portfolio - Statistics

Quantitative Portfolio - Statistics

Quantitative Portfolio - Statistics

Quantitative Portfolio - Statistics

Portfolio Comparison

Polar Star Products and S&P500

Combine Discretionary and Systematic Portfolios

Polar Star Multi Strategy as enhancement to S&P500

Portfolio Comparison

Combining a

• discretionary and
• systematic approach

to investing in commodities we create a product with

• positive expected return which is
• uncorrelated to equities

that gives superior risk adjusted returns.